The signature is still yours
A quiet day for regulators and a busy one for the tools: a 25-seat Copilot trial opens to small firms, capable models sit at pennies per million tokens, and the phrase doing the rounds is about what happens when nobody reads the output before passing it on.
- 01
Microsoft opens a 25-user Copilot trial for firms under 300 people
From 1 August, Microsoft partners can transact a new 25-user, 30-day Microsoft 365 Copilot Business trial through CSP, aimed at organisations with fewer than 300 employees. It comes wrapped in a partner programme called Copilot in 30, with setup guides and a success planner tool due mid-August. Practical point: if you've been quoted per-seat Copilot licences and stalled, your IT provider now has a structured way to let you try it before committing.
Microsoft Partner Center announcements - 02
A million-token model at 14 cents in, 28 cents out
DeepSeek's published pricing puts V4-Flash at $0.14 per million input tokens and $0.28 out, with V4-Pro at $0.435 and $0.87, both with a one-million-token context. The same page flags that a peak/off-peak policy is coming, doubling prices between 09:00–12:00 and 14:00–18:00 Beijing time, with no start date given. Worth knowing if you or your developer priced an automation against these numbers: the headline rate isn't promised to stay flat all day.
DeepSeek API docs - 03
Google Meet notes now capture what was on screen
Rolling out from 3 August, the 'Take notes for me' feature can drop screenshots of presented content into the notes document, so a drawing or programme extract shown in a call ends up in the write-up rather than only in the transcript. It's on Business Standard and Plus, Enterprise Standard and Plus, and AI Pro for Education, and rollout is gradual. It isn't on by default: an admin decides whether screenshots are always allowed or only during a recording, and the presenter is notified and can turn it off.
Google Workspace Updates - 04
EU code signatories asked to be open about their crawlers
The Commission published the readout of the fourth GPAI Code of Practice taskforce meeting, covering the safety and copyright chapters. On copyright, signatories commit to letting rightsholders find out which web crawlers they run and how those crawlers read robots.txt, and to publish how they identify a reservation of rights while crawling. If you publish anything on your own site — case studies, drawings, photos — that's the mechanism by which saying no is meant to work.
European Commission - 05
'Don't be a meat proxy' — the phrase of the week
Gruhn's post names a habit a lot of people will recognise: someone pastes a question into a model, then pastes the answer straight into Slack or a review comment without reading it, and does the same in reverse with the reply. His line is that the recipient could have asked the model themselves, faster and with better context. Simon Willison picked it up the same day and added the fix: read it, understand it, check it, then write your own words.
Niklas Gruhn - 06
Cloudflare squeezes the big open models to serve them cheaper
Cloudflare wrote up how it now serves Kimi K2.6 and GLM 5.2 on Workers AI: an 8-bit KV cache roughly doubles the context it can hold, and compressing GLM's weights to 4-bit integers cut the checkpoint from about 705GB to 421GB. No price change was announced, and they claim no accuracy loss. It's the plumbing behind the trend that actually reaches you — the cost of running a serious open model keeps falling, which is why the quotes you get for AI work keep falling too.
Cloudflare
The phrase I'd keep from today is Gruhn's. A meat proxy is someone who passes AI output along without reading it, in both directions, and adds nothing but a delay.
It reads like a developer complaint, but it's a trades problem with a longer history. On site I signed things. A method statement, a snag list, a sheet saying a flat was ready for the next trade. The document was only worth anything because a person who understood the work had checked it and put their name to it. Nobody ever asked whether I'd typed it myself.
That's the part that survives. Let the model write the RAMS, the quote, the report to the main contractor — genuinely, let it, that's hours back. But the name at the bottom is a claim that you read it and it's true. If a supervisor forwards a generated method statement unread, they haven't saved an hour, they've moved the checking onto whoever opens it next, and that person didn't agree to it.
The tools in today's issue all point the same way: cheaper to run, easier to trial, more of your week automatable. None of them change who's liable when the duct goes in on the wrong route because the drawing reference in the paperwork was invented.
The odd thing is that a firm too small to have a compliance department is better placed here, not worse. The person signing is usually the person who did the work.
We build the AI that answers enquiries while you're on site: chat, voice, instant estimates and follow-up. See how it works or price it in two minutes.